OpenAI proudly announces that GPT-5.6 Luna text dialogue is now freely available worldwide to all users, becoming the default model for Free and Go tiers, and introducing the new “Think” button for the free version for the first time. Simultaneously, the most powerful model, GPT-5.6 Sol, has been fully upgraded, with a new Thought Depth slider added to the ChatGPT interface, unifying Instant and Thinking modes and allowing users to freely adjust the AI’s reasoning depth. In rigorous testing across finance, healthcare, and law, GPT-5.6 Sol reduced error rates by 68% compared to GPT-5.5, while Luna reduced errors by 62%. Additionally, insiders report that OpenAI will launch its next flagship model, Astra, next week; its internal codename “mewfour” has reached release candidate status, with an expected parameter scale of 7–10 trillion, designed for long-duration multi-agent collaboration.Article author and source: AI New Era
No way?
Even GPT-5.6 is now completely free for everyone!
This morning, OpenAI officially announced that text conversations with GPT-5.6 Luna are now unlimited.
One billion people worldwide can use it directly for free.
Meanwhile, GPT-5.6 Sol has received a major upgrade—
The ChatGPT interface now features a new slider that unifies Instant and Thinking modes, allowing Sol to determine how long to think about each question.
GPT-5.6 is truly free now
On July 9, OpenAI unveiled the entire GPT-5.6 suite at once, and Sol decisively outperformed the top Fable 5.
Less than a month ago, GPT-5.6 Luna was opened for free access.
Unlimited sounds great, but there's one thing we need to clarify.
Luna is the smallest in the GPT-5.6 family, designed for speed and affordability, alongside Terra and Sol.
Starting today, the "Mini" Luna model will become the default model for Free and Go, replacing GPT-5.5 directly.
In addition, OpenAI has added a new "Think" button to the free version of ChatGPT—
Encounter a difficult question? Just click once to let the model think longer before responding—this is the first time the free tier can actively invoke higher reasoning power.
However, this free, unlimited offering applies only to "text chat."
File uploads, image generation, voice features—limits remain the same and will still hit the cap.
Super-sized SOL, fully upgraded
OpenAI is offering a major promotion, along with an upgrade to the most powerful GPT-5.6 Sol.
Previously in ChatGPT, the quick response of Instant and the slow, thoughtful Thinking were like two distinct personalities:
The tone is different, the format is different—it’s like switching to a different model.
It is now uniformly managed by Sol, allowing seamless speed control—the only difference being the length of "thinking time," with no more sense of fragmentation.
The "Five-Level Thinking Slider" feature, previously exclusive to ChatGPT Work, is now directly available on the standard chat interface!
Everyone can freely adjust the depth of AI thinking.
The most obvious change in the new Sol is that it "talks less," providing more concise and to-the-point responses.
In its official blog, OpenAI gave an example: "Will I get soaked cycling from Mission to the beach after work?"
The old version of Instant listed a long string: rainfall, wind speed, temperature, sea fog, beach danger warnings, plus the note, “You might feel sticky wearing a cotton T-shirt.”
The new Sol starts with the conclusion: you won’t get wet; the real issue is the headwind—10 to 20 mph from the west—so just bring a light wind-resistant jacket.
Then, when asked "I'm leaving at 5:30," the old version replayed the entire forecast, while the new version only updates the conclusion.
Swipe up or down to view
Swipe up or down to view
This is the entire essence of the GPT-5.6 Sol update: replace "comprehensive" with "focus on the key points."
Error rate drops by 68%
There’s one more important thing: the factual accuracy of the GPT-5.6 “suite” has significantly improved this time.
The evaluation criteria for this round are very strict.
They conducted internal evaluations on issues in the three fields most intolerant of misinformation: finance, healthcare, and law.
If a single answer contains even one factual error, the entire answer is considered incorrect.
As a result, GPT-5.6 Sol has a 68% lower error rate than GPT-5.5 Instant, and Luna has a 62% lower error rate.
GPT-6 is coming next week?
Free? That's just the appetizer.
Today, industry-famous leaker Leo revealed exclusively: OpenAI is preparing to launch its next flagship model, Astra, next week.
Astra is the largest model OpenAI has trained since GPT-4.5, representing a completely new pretraining effort. The latest internal checkpoint, codenamed "mewfour," is already a Release Candidate.
RC stands for the final version before release!
In fact, there has long been a trail of unmistakable signs.
On July 30, a preview video from OpenAI was quickly deleted. A netizen quickly took a screenshot: the screen briefly flashed the word "mewthree."
This week, further leaks revealed that mewfour is already in internal testing.
Counting from Mew to Mewtwo, OpenAI’s “Mewtwo” has quietly evolved into its fourth form.
More interestingly, the official announcement was already right out in the open.
On August 1, OpenAI published a mathematics blog stating that its model solved 10 open problems that had remained unsolved for over a decade.
No one expected that the official announcement of the next-generation model was hidden in this blog post:
These results were achieved with an internal version of Astra—our next major model.
So, what exactly is this Astra?
Based on current leaks, its positioning is long-duration, multi-agent collaboration—where a group of AIs work together for hours or even days to solve a complex problem.
Rumors suggest there are two calibers.
Previously, media reports suggested it was twice as large as GPT-5.6; renowned developer Haider estimated that GPT-4.5 has around 5 trillion parameters, while Astra may reach 7–10 trillion. In short, it’s a behemoth.
However, with stronger infrastructure and optimizations, along with a large amount of computing power coming online this year, OpenAI may run this massive model at a lower service cost than Anthropic’s MythoMax-5.
In Haider’s view, OpenAI’s post-training is the strongest in the industry, and that’s how GPT-5.5’s capabilities were extracted; its only disadvantage compared to Mythos is its pre-training foundation.
This time, the weakness was directly turned into the largest in history.
The strongest emphasis on education, paired with the largest foundation—that’s the entire logic behind Astra’s potential breakthrough.
In other words, this would be a Mythos/Fable-level leap for OpenAI.
And all of this is done for one thing only—to dethrone Anthropic's Fable 5.
As for whether Astra is a generational leap or just slightly ahead,
The answer may be revealed within a week.